For CISOs and security teams
Find the shadow AI. Then give it somewhere safe to go.
Sonsara shows you every AI tool touching company systems, on whose account and with which data, from the logs you already keep. Then it moves the work worth keeping onto governed company agents, with the audit trail and evidence your ISO 27001, ISO 42001 and SOC 2 reviews will ask for.
IBM, Cost of a Data Breach, 2025
IBM, Cost of a Data Breach, 2025
The exposure
The AI you didn't approve is already reading company data.
Personal accounts, OAuth consents to AI apps, tokens pasted into prompts, browser agents acting as employees. Each one is a data flow nobody signed off, and most of them are invisible to the tools you have.
01 · Data
Personal accounts on company work.
Supplier details, customer notes and draft financials going into AI accounts outside any contract or data agreement, and staying in their history after the person leaves.
02 · Identity
Agents acting as employees.
People connect AI apps to mail, drive and code with OAuth, or hand an agent a personal access token. The agent then acts as them, with all their access, and nobody reviewed the grant.
03 · Evidence
No record to show an auditor.
When an assessor, a customer's security questionnaire or your insurer asks how AI is used and controlled, there's nothing to hand over.
Detection, privacy-first
Metadata, not messages.
Detection starts with read-only access to your identity provider and work tools' audit logs. One app consent or admin key per system, nothing installed on laptops, first findings in days.
- Reads which system, which client, which app, timing and volume
- Never reads prompts, model responses or the contents of files and messages
- People are pseudonymised; re-identifying someone is an audited admin action
- Optional deeper layers, when you want them: existing DNS, gateway and SIEM logs, or a device check through your own MDM
- Consents to AI appsIdentity provider · app id, scopes, dateSignal
- New API tokensSaaS audit logs · created, used bySignal
- Programmatic clientsAudit logs · client type, user agentSignal
- Machine-paced accessAudit logs · bursts, even intervalsSignal
- Prompts and contentNever collectedNever
Controls on company agents
Say yes to AI, on terms you set.
Blocking a tool sends the work somewhere you can see even less. Company agents keep the time people were saving and give you the controls: least privilege, set regions, approved models, and a record of everything.
Guardrails
Block, tag or log.
Policies block email outside the company, tag personal data in outputs, and log every tool call. Risky actions wait for a named approver.
Data & models
In region, approved models only.
Agents run in a set data region, such as EU · Dublin, on approved models. Customer data is read where it lives and never copied.
Credentials
Company-held and scoped.
Connectors sit on the agent's account, not an employee's, with only the access the job needs. When someone leaves, nothing runs on their login.
Access
Owners, editors and users.
You decide who can use an agent, who can change it, and who approves a new version before it goes live. Single sign-on throughout.
Change
Tested before it ships.
Every change is a new version, test-run and checked against its evals. The current version keeps running until an owner approves the new one.
Watched
Broken ones flagged.
Agents are watched while they run. A broken one is marked down and its owners are asked to fix it.
Audit-ready
Evidence for ISO 27001, ISO 42001 and SOC 2, kept as agents run.
The activity log records each run, tool call, policy block, change and access grant, with who and which data. Here's where that record supports the control areas auditors look at for AI. Export it for an assessor, a customer questionnaire or a cyber insurance review.
FAQ
What security teams ask us.
What exactly do you collect about employees?+
Metadata from systems you already log: which system, which client or app, when and how often, plus consents and token creation. People appear as pseudonyms. We never collect prompts, responses, file contents or browsing history beyond visits to known AI services.
Does detection need an agent on every device?+
No. It starts with read-only access to your identity provider and work tools. A device check through your own MDM is an optional later layer, if you want to tell company accounts from personal ones on managed laptops.
Is Sonsara certified for SOC 2 or ISO?+
What Sonsara gives you is evidence for your own audits: the activity log, version approvals, access records and an inventory of AI tools and agents. Your auditor decides what they accept.
How does it fit with our CASB, DLP or SSE?+
Alongside them. Your existing logs can feed detection, and Sonsara adds what comes after a finding: a governed company agent for the job, so the work has somewhere safe to move to.
Can we see what an agent did, after the fact?+
Yes. Every run is in the activity log with the tools it called, the data it touched, its outcome, and who approved the version that ran. Run history is kept for 30 days, and you can export it to keep it for longer.
Keep reading
The detail behind the controls.
Book a demo
See the AI work already happening in your company.
Read-only access, nothing to install, first findings in days.